agentleFS
dsebastiendsebastien/ai-skill-scholar/skills/literature-review/SKILL.md

literature-review

Skill16 starsChanged 7 months agoMITView on GitHub

Orchestrate a two-pass literature review for a research question. Phase 1 casts a wide net via scholar-search (and optionally arxiv-search), phase 2 the agent screens candidates by title+abstract, phase 3 the shortlist gets full-text analysis, phase 4 the agent writes a structured synthesis. Persistent session state lives in ./literature-reviews/<slug>/. Use when the user says "literature review", "survey the field", "research question", "what does the literature say about", or "systematic search".

89 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use pdf-explore, paper-narrative and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • JimLiuSkill

    pdf-explore

    Use this skill when the user has attached a PDF, paper, report, or other document and the answer needs content from more than one place in it: summarize the methods or any other section, compare sections, find where a topic is discussed, read a value or label off a figure or chart, or find/list/extract every instance of something across the whole document (datasets, benchmarks, citations, figures, table rows, accession numbers — including appendices). Skip it only for a single lookup of 1–4 pages quoted in your very next response — `read_file(pages=[...])` attaches pages as images that are dropped from context after one turn, so multi-section answers end up re-reading the same ranges repeatedly. Parses the PDF once in the Python kernel: `pdf_pages` (pages as persistent text), `pdf_outline` (TOC), `pdf_scan` (rank pages by relevance), `pdf_map`/`pdf_extract` (per-page summary / structured fields via parallel haiku calls). For PDF creation/manipulation, use reportlab/pypdf directly.

    Also in 9 projects that use this

  • JimLiuSkill

    paper-narrative

    Judge and reshape the STORY a paper's figures tell. Input is the work itself — manuscript (or abstract) + figure deck — no hand-written brief. `derive_paper_brief(abstract, captions)` extracts pitch/vision/per-figure-claims; a handling-editor reviewer on the full deck returns hook_verdict (would Fig 1 make me send this for review?), arc (hook→mechanism→evidence→application), figure_moves (panels in the wrong figure), missing_panels (concrete analyses to RUN), kill_list, and boldest_defensible_fig1. Hands per-figure claims to `figure-composer`. Load when writing or revising a paper.

    Also in 5 projects that use this

  • Norman-burySkill

    brainstorming-research

    Use before writing any paper - clarifies paper type, discipline, topic, methodology, and chapter structure through natural dialogue

    Also in 4 projects that use this

  • lingzhi227Skill

    latex-formatting

    Handle LaTeX formatting, templates, and styling for academic papers. Set up conference templates (ICML, ICLR, NeurIPS, AAAI, ACL), fix formatting issues, manage packages, and ensure venue-specific compliance. Use when the user needs to set up a paper template, fix LaTeX formatting, or prepare for submission.

    Also in 4 projects that use this

  • lingzhi227Skill

    novelty-assessment

    Assess research idea novelty through systematic literature search. Multi-round search-evaluate loops with harsh critic persona. Binary novel/not-novel decision with justification. Use before committing to a research direction.

    Also in 4 projects that use this

  • +6 more from JimLiu
  • +1 more from lingzhi227

The file

We split this skill into the rules it gives an agent

5 rules. 0 are common, 0 are distinctive, and 5 are very specific to this project.

  • 0

    Common

    Most of the 69 edited versions we read keep these.

  • 0

    Distinctive

    This author's own choices. Worth a look.

  • 5

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Default limit 50. Use --year to scope recency, --min-citations to cut noise.

    line 74
  2. Very specific:

    Read candidates.json. For each paper, decide: include or exclude.

    line 79
  3. Very specific:

    Write final.md with this structure:

    line 101
  4. Very specific:

    Never fabricate citations or TLDRs: If a field is missing in shortlist.json, mark it as such in the synthesis.

    line 130
  5. Very specific:

    Session state is append-only by convention: If you need to revise, prefer creating a new session over editing state.json in place.

    line 132

“Common” is measured: kept by at least 6 in 10 of the 69 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Literature Review

Orchestrate a structured two-pass review. The script drives state management and data fetching; the agent drives the judgment calls (screening, synthesis). Each review is a persistent session — you can pause, resume, and revisit.

Phases

1. init      Create session, capture the research question
2. search    Wide net across Semantic Scholar (+ arXiv optional), dedup
3. screen    Agent reads titles/abstracts/tldrs, marks keepers
4. fetch     Script produces a fetch plan (arxiv-analyze or PDF URL per paper)
5. [read]    Agent reads each shortlisted paper via the indicated strategy
6. [synth]   Agent writes the review: established findings, tensions, gaps

Steps 5 and 6 are agent-driven (no new subcommand); the script's job is to set up the material.

Usage

# 1. Start a session
python3 scripts/literature_review.py init \
    "what makes sparse autoencoders interpretable in practice" \
    [--out-dir /path/to/reviews]

# Returns: {"session": "./literature-reviews/what-makes-sparse-..", "slug": "..."}

# 2. Cast a wide net
python3 scripts/literature_review.py search <session-dir> \
    --limit 50 --year 2023-2026 --min-citations 5 \
    [--venue ICLR,NeurIPS] [--include-arxiv]

# 3. Agent reads candidates.json, picks keepers
python3 scripts/literature_review.py screen <session-dir> \
    --include "a3ec0b75...,2501.11120,10.48550/arXiv.2401.00032" \
    [--reasons-file keep-reasons.json]

# 4. Get the fetch plan
python3 scripts/literature_review.py fetch <session-dir>

# 5-6. Agent reads each paper and writes ./literature-reviews/<slug>/final.md

# Status check at any time
python3 scripts/literature_review.py status <session-dir>
python3 scripts/literature_review.py list-sessions [--out-dir /path]

Session layout

./literature-reviews/<slug>/
├── state.json         # phase, question, counters
├── candidates.json    # phase-1 output: full search results
├── shortlist.json     # phase-2 output: papers the agent kept + reasons
├── fetch_plan.json    # phase-3 output: per-paper read strategy
└── final.md           # phase-4 output: agent writes this

Everything is JSON except final.md. Sessions can be re-run (rerun search to refresh, redo screen to revise shortlist). Delete the session dir to start over.

Workflow

Phase 1: init

Capture the research question clearly. Good questions are specific: "how does X differ from Y under condition Z" beats "X and Y". If the question is vague, prompt the user to sharpen it before running init.

Phase 2: search

Default limit 50. Use --year to scope recency, --min-citations to cut noise. Add --include-arxiv if arxiv-search is installed and you want preprint coverage the scholar index may lag behind on.

Script dedups by arxiv_id, doi, and lowercased title.

Phase 3: screen (the agent does this)

Read candidates.json. For each paper, decide: include or exclude. Use the first 1-2 sentences of the abstract for fast scanning; drill into the full abstract only for ambiguous calls.

Explicit inclusion criteria help consistency:

  • Does the paper address the question directly?
  • Is the method/claim specific enough to learn from?
  • Is the venue/year/citation count above threshold?

Invoke screen with the comma-separated IDs of keepers. Optionally pass --reasons-file mapping each included ID to a one-line reason — these flow into shortlist.json and into the final synthesis.

Target shortlist size: 10-25 papers. Fewer and you're missing coverage; more and synthesis gets unwieldy.

Phase 4: fetch

Script produces fetch_plan.json: for each shortlisted paper, the best way to read it:

  • arxiv-analyze — when arxiv_id is present (preferred: tiered markdown → HTML → TeX fallback)
  • open-access-pdf — direct PDF URL
  • abstract-only — no full-text available; agent decides whether to keep

Phase 5: read (agent-driven)

For each paper in the fetch plan, pull the full text using the indicated strategy. Extract: claim, method, result, limitation. Keep notes in memory or in scratch files inside the session dir.

Phase 6: synthesize (agent-driven)

Write final.md with this structure:

# Literature Review: <question>

## TL;DR
One paragraph. What does the literature say?

## Established findings
Claims multiple papers converge on, with citations.

## Active debates
Claims papers disagree about. Who holds each position, and on what grounds.

## Methodological patterns
Dominant approaches and their trade-offs.

## Gaps
What's unstudied, understudied, or inconclusive.

## Most important paper to read next
One paper. Why.

## Appendix: included papers
Table: title | authors | year | venue | one-line takeaway | full-text status

Hard rules

  • Never fabricate citations or TLDRs. If a field is missing in shortlist.json, mark it as such in the synthesis.
  • Two-pass screening is not optional. Skipping the screen phase and going straight to full-text reads for 50 papers is a token disaster and a quality disaster (signal-to-noise).
  • Session state is append-only by convention. If you need to revise, prefer creating a new session over editing state.json in place.
  • Target ~10-25 shortlist papers. Fewer misses coverage; more is unwieldy.
  • Note what you couldn't read. abstract-only papers are mentioned in the review with a note that full-text was unavailable — don't pretend you read them.

Token budget

PhaseTokens (rough)
search (50 results with abstracts)15K-30K
screen (agent skim)15K-30K in, minimal out
fetch plan2K
read (15 papers via arxiv-analyze md tier)150K-300K
synthesis5K-15K

Total: a real review burns 200K-400K tokens. Plan accordingly — split phases across sessions if your context budget is tight.

The One Thing

End every status report with the single most impactful next action for the current phase — not a status summary. Example: "Phase: screened (12 shortlisted). Next: run fetch, then start with paper 2501.11120 — it's cited by 6 others in your shortlist, so reading it first will anchor the rest."

Requirements

  • Python 3.11+ (stdlib only)
  • scholar-search skill installed as a sibling (same skills/ directory)
  • Optional: arxiv-search and arxiv-analyze skills for arxiv integration (install from https://github.com/dsebastien/ai-skill-arxiv)
  • Optional: OPENALEX_EMAIL env var for polite-pool access (no API key needed; OpenAlex is free)

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.